| 你們好，大叔謝謝你們訂閱，也為訂閱者提供 📢公開討論群📢：Link 歡迎加入大家一起互動。
Industry: Technology, Media, and Telecom (TMT) / AI Infrastructure / Data Centers Coverage: US Public Equities After Generative AI emerged from the labs, the global computing market was sucked into a hardware famine. NVIDIA H100, B200, and GB200 GPUs are now as scarce as wartime strategic resources. Most interestingly, the traditional cloud giants—AWS, Azure, and Google Cloud—burdened by heavy architectures, slow processes, and complex supply chains, initially lost the footrace in this GPU shopping spree to some nimble “Neoclouds.” Based on a key report by McKinsey & Company (The evolution of neoclouds and their next moves), this emerging sector has been deconstructed to analyze its investment value. McKinsey’s analysis points out that while Neoclouds played a critical role during the GPU shortage, the “Bare Metal as a Service” (BMaaS) business model is economically fragile. Simply put, McKinsey has thoroughly analyzed the entire context of Neoclouds, and as an investor, you cannot afford to ignore it. The core conclusion is direct: Neoclouds are heroes during the GPU shortage, but they cannot survive long-term relying solely on the “Bare Metal Rental” model. Why? Because BMaaS has three fatal inherent flaws: Hardware Depreciation is like a Fire Sale: NVIDIA releases a new generation of GPUs every year. The H100 you buy today might be crushed by the B200 or Rubin in three years. Recouping the rental cost becomes a life-or-death line. Off-the-Charts Capital Pressure: If you don’t buy the latest GPUs, you lose. If you buy too many and recoup costs too slowly, you also lose. Erosion of Pricing Power: Once the supply chains of AWS and Azure catch up, the “cheap and fast” advantage of Neoclouds will shrink.
That is why McKinsey says: This industry is now stuck at a fork in the road—“Evolve or Die.” To understand this round of AI infrastructure revolution, you must first understand why Neoclouds appeared, why they were hyped to the sky, and why McKinsey says they must evolve. Neoclouds didn’t grow out of technological innovation; they were squeezed out of the “Global GPU Famine.” Force 1: Severe Shortage of High-End Compute: Models like GPT-4, Claude 3, Llama, and Gemini are getting bigger. Training takes thousands of GPUs; inference requires massive compute. Demand exploded, supply couldn’t keep up, and GPUs became gold-tier assets. Force 2: NVIDIA Didn’t Want to be Tied to the Big Three: AWS, Azure, and Google were eating up too much supply. To reduce concentration, NVIDIA started nurturing new channels. Neoclouds like CoreWeave, Lambda, and Nebius received “priority allocation.”
This was the Phase 1 Value Proposition of Neoclouds: “We are fast and cheap.” For teams desperate to train models, this was a lifeline. But McKinsey threw cold water on this: This arbitrage model won’t last long. The reason is just two words: Hardware Depreciation. What is the basic business of a Neocloud? It is “renting out GPUs.” On the surface, this model looks delicious: Gross margins of 55%–65%, high rental ceilings, and exploding demand. But McKinsey revealed the underlying truth: These are pre-depreciation margins. After depreciation, profits are paper-thin. Depreciation Death Spiral: NVIDIA now iterates annually: Hopper → Blackwell → Rubin → (New one next year). This means asset values shrink rapidly. An H100 might lose over 50% of its rental value in three or four years. Massive Annual Capex: To keep up with performance, you have to keep buying new GPUs. The old GPUs haven’t paid for themselves yet, and the new ones are already here. Profit is an Illusion: After factoring in depreciation: If revenue is $100, depreciation eats $54, leaving an EBIT of only 8%.
McKinsey warns: Today’s situation is very similar to the “Cloud 1.0 Era” after the 2000 dot-com bubble. Back then, many hosting providers rose on cheap prices and fast deployment, but were eventually washed out by AWS and Azure because they lacked a software moat. If today’s GPU renters don’t evolve, they will follow the exact same path: Get acquired or get eliminated. McKinsey is blunt: There are only three ways for Neoclouds to survive. Move Up the Stack: The core: Shift from “renting hardware” to “providing services.” Includes: AI Native Orchestration, Inference Platforms, Managed Services, Custom Training Platforms. These offer high margins, stickiness, and valuations. Companies like CoreWeave and Nebius have the highest win rate here.
Niche Markets: Don’t fight a price war with AWS; avoid them, bypass them, outperform them in specific areas. Potential niches: Sovereign Compute, Defense & Government (the main reason “Jensen’s godson” companies touch the US government), High-security industries like life sciences. These customers don’t chase the lowest price; they chase compliance, data isolation, and geopolitical security. Oracle and Palantir are becoming gatekeepers here.
Consolidation:
McKinsey is clear: The opportunity for Neoclouds is huge, but so is the risk. The survivors will be those evolving toward software, niches, or consolidation. Pure GPU rental won’t last. Investors shouldn’t look at exploding revenue or GPU count, but rather: Is this company on the “Right Path to Survival”? (Important enough to say three times). Not every Neocloud is investable. The real future belongs to companies that are not just renting GPUs but becoming “AI Platforms,” and are viewed by NVIDIA as strategic partners. Status: The definer of the Neocloud sector. Not a rental shop, but building “AI Infrastructure.” IPO on Nasdaq in March 2025 at $40/share. Market cap topped $36B by Nov 2025. Top-tier NVIDIA partner and a key supplier for Microsoft/OpenAI. Moving Up the Stack: Unlike traditional providers offering VMs, CoreWeave offers “Kubernetes Native Cloud for AI Training.” Their Orchestration Layer offers low latency and high distributed efficiency—a true software moat. 2025 Acquisition of Core Scientific: A critical move. From “renting a room” to “building the community.” They acquired 1.3 GW power capacity and data centers. This vertical integration (Power + DC + GPU + Software) avoids the pure BMaaS depreciation trap. NVIDIA’s Endorsement: NVIDIA is both a shareholder and a partner, allocating the first batches of Blackwell/Rubin chips to CoreWeave as a strategic pawn against AWS/Google’s in-house chips. Uncle’s Reality Check: Q3 2025 revenue grew 133% YoY. However, customer concentration (Microsoft/OpenAI) is the biggest risk. If Microsoft moves workloads to its own Maia chips, it leaves a gap. Aggressive 6-year depreciation poses future asset impairment risks.
Status: Born from Yandex’s international arm, now HQ’d in Amsterdam and listed on Nasdaq. Positions itself as an “AI Native” infrastructure company. Engineering Culture: Inherited top-tier talent from Yandex. They design their own server racks, cooling, and cloud platforms. Conservative Financials: Uses a 4-year straight-line depreciation for GPUs (vs. CoreWeave’s 6 years). This is conservative and aligns with Moore’s Law, resulting in higher “Quality of Earnings” and lower risk of future write-downs. Sovereign Strategy: Building clusters in Europe (Paris) and US (Kansas). Strong compliance advantage (GDPR) for European enterprise/government contracts. Investment View: A strong rival to CoreWeave but with a different style. CoreWeave is “Vertical Integration + Platform”; Nebius is “Engineering Depth + Compliance + Quality Earnings.” Lower valuation multiples make it suitable for investors preferring technical purity and moderate risk.
Logic: Power is the speed limit for AI. Whoever holds powered land is King. McKinsey confirms: The next limit to AI growth is Power. US data center hubs (Virginia, Dallas, Phoenix) have 5-7 year wait times for power. The unexpected winners? Bitcoin Miners. They possess “Powered Shells”—ready-to-use, high-power density sites. This is the HPC Pivot. Strategy: Building “AI Factories,” not traditional saunas. Focused on large-capacity facilities for H100/Blackwell clusters. Core Competency: Ultra-high density (40–100kW+ per rack) vs standard 15kW. Air cooling + Liquid cooling ready. Endorsement: Signed a massive 400 MW lease with CoreWeave. This is essentially a REIT model: stable rent, independent of crypto prices. A CoreWeave contract is an indirect NVIDIA endorsement. Clean Balance Sheet: Divested legacy blockchain hosting. Re-rated from “Crypto” to “AI Infra.”
Strategy: Hybrid. Use Bitcoin mining cash flow to buy GPUs and build HPC clusters. Data: AI Cloud revenue is small but growing fast (33% QoQ). Hardware margins are 97%. Bold Move: Buying GPUs directly (H100/Blackwell) instead of just hosting. This accepts depreciation risk but gains higher margins and platform potential (”Move up the stack”). Asset: 3 GW power reserve, mostly renewable (Hydro, Wind, Solar). A huge plus for ESG-conscious tech giants and governments.
Story: Best case of infrastructure re-rating. Post-bankruptcy, signed with CoreWeave, proving its mines could upgrade to HPC. Logic: With CoreWeave announcing a full acquisition, CORZ stock became an M&A arbitrage tool. Investors are indirectly holding future CoreWeave equity.
Others to Watch: Hut 8 ($HUT), TeraWulf ($WULF). Uncle’s Summary: The logic is clear. GPU precious → Power precious → Miners valuable. Look for: Massive powered land, High-density support, AI customers (ideally CoreWeave), and a formal shift to “AI Factory Operator.” Logic: Neoclouds need more than just GPUs. Efficiency depends on connection, transmission, memory, and storage. These companies are the picks and shovels—the nervous system and blood vessels. Role: The “Translator” between GPU and CPU. Importance: In the PCIe 5.0/6.0 era, signals degrade over centimeters. Astera’s Aries Retimer reconstructs signals. It is standard in NVIDIA HGX/MGX designs. COSMOS Software: A “Maintenance Lifesaver” for Neoclouds. Monitors PCIe link health, predicts failures, and reduces operational costs for massive clusters. Growth: Multi-product line (Scorpio switches, Leo CXL controllers). As servers get stronger, Astera becomes more vital.
Role: The “Highway” between servers. AEC (Active Electrical Cable): Contains DSP chips. Cheaper, thinner, and 50% lower power than optical fiber. ZeroFlap: Connection technology designed for AI to prevent training interruptions (re-runs are expensive). Value: Energy efficiency translates directly to money. Saving 50% power on cables saves millions for a mega-cluster.
Arista Networks ($ANET): High-performance Ethernet switches. The “Network Arsenal” for Neoclouds who don’t want to be locked into NVIDIA’s InfiniBand. Pure Storage ($PSTG): “Glucose supply” for GPUs. If storage is slow, GPUs wait. CoreWeave has partnered with them. FlashBlade is essential, not optional.
Logic: AI is moving from a tech war to a geopolitical war. The winners are those who guarantee “Data doesn’t leave, System isn’t controlled.” Sovereign Compute is the only niche where Neoclouds can avoid direct assassination by AWS/Azure. Distributed Cloud: OCI Dedicated Region. Puts the entire public cloud stack inside the customer’s data center (or bunker, or submarine). EU Sovereign Cloud: Operated by EU entities, compliant with EU laws, isolated from the US Cloud Act. The “Second Supplier”: OpenAI and Microsoft use OCI for training/inference redundancy. Low latency RDMA networking + High efficiency.
AIP: Runs LLMs in disconnected (air-gapped) military bases. IL6 Certification: Highest US military classification. A barrier to entry that standard companies cannot touch. Alliance with Oracle: Oracle (Hardware/Architecture) + Palantir (OS/Governance) = The strongest solution for government clients.
Uncle’s Summary: Winners here have three traits: Data absolute residency, capability to run AI offline, and Defense-grade endorsements (IL6, NATO). Logic: In a gold rush, sell water. These companies don’t bear the depreciation risk of GPUs but ride the Capex wave. Datadog ($DDOG): The “Observability Brain.” AI models hallucinate, overheat, and lag. DDOG’s LLM Observability is the health tracker. AI customers are their new growth engine. Arista ($ANET): The Anti-NVIDIA Network Alliance. 50% of training performance is the network. Pure Storage ($PSTG): Feeding the GPU’s stomach. Super Micro ($SMCI): Note: Accounting flag acknowledged. But they act as the “Rapid Deployment Force” via modular servers for Tier 2/3 Neoclouds.
Depreciation Cliff: If utilization drops below 80%, depreciation eats all profit. H100 values drop like fresh produce. Watch: Depreciation period (shorter is better). Customer Concentration: If Microsoft leaves CoreWeave, it’s a disaster. Strategy: Diversify (Oracle, APLD). Cost of Capital: High interest rates crush high-leverage Capex models. Watch: Debt/EBITDA, Cash Flow. Geopolitical Risk: Chip bans and data sovereignty. Strategy: Buy policy beneficiaries (Oracle, Palantir).
“The AI Infrastructure Battle Royale” Left Hand (Offense): High Growth / High Risk CoreWeave ($CRWV): The King. Vertical integration. Risk: Concentration. Nebius ($NBIS): Quality engineering, conservative accounting, sovereign advantage. Risk: Smaller scale.
Right Hand (Defense): Hard Assets / Tools Astera Labs ($ALAB): Signal clearing monopoly. Credo ($CRDO): Cabling revolution. Suitable for: Investors wanting participation with lower volatility.
Middle (Hedge): Sovereign Moats Summary: You will see GPU renters collapse, crushed by capital costs and depreciation. The survivors will possess: Power + Software + Sovereignty. If you have all three, you are the next Super Hyperscaler. Appendix: Key Ticker Symbols Neoclouds: $CRWV, $NBIS Miners/Infra: $APLD, $CORZ, $IREN, $HUT, $WULF, $CIFR Sovereign/Defense: $ORCL, $PLTR, $GOOGL, $AI Enablers: $ANET, $PSTG, $DDOG, $SMCI, $VRT, $ALAB, $CRDO
【Disclaimer】 The content provided on this site represents the author’s professional analysis and personal observations regarding companies, industry trends, and financial conditions as of a specific date. It is intended for informational and reference purposes only and does not constitute any solicitation, inducement, recommendation to buy or sell, investment advice, or guarantee of profit or loss. LINK: The evolution of neoclouds and their next moves by McKinsey & Company CoreWeave, Inc. (CRWV) IPO - NASDAQ.com, accessed November 30, 2025, https://www.nasdaq.com/market-activity/ipos/overview?dealId=1079716-113203 CoreWeave Announces Pricing of Initial Public Offering, accessed November 30, 2025, https://investors.coreweave.com/news/news-details/2025/CoreWeave-Announces-Pricing-of-Initial-Public-Offering/default.aspx Stock Info - CoreWeave - Investor Relations, accessed November 30, 2025, https://investors.coreweave.com/stock-info/default.aspx CoreWeave, Inc. (CRWV) IPO | IPO Calendar - Click Capital Markets, accessed November 30, 2025, https://click.markets/symbol/crwv CoreWeave to Acquire Core Scientific, accessed November 30, 2025, https://investors.corescientific.com/news-events/press-releases/detail/119/coreweave-to-acquire-core-scientific CoreWeave Inc (CRWV) Stock Price & News - Google Finance, accessed November 30, 2025, https://www.google.com/finance/quote/CRWV:NASDAQ GPU Depreciation: CoreWeave vs. Nebius - Tech News - Bizety, accessed November 30, 2025, https://bizety.com/2025/09/23/gpu-depreciation-coreweave-vs-nebius/ Nebius to build first NVIDIA GPU cluster in US; expands offices to serve customers across the country, accessed November 30, 2025, https://nebius.com/newsroom/nebius-expands-in-us-with-first-gpu-cluster-in-kansas-city-offices-in-san-francisco-dallas-and-new-york Nebius to invest more than USD 1 billion to build AI infrastructure in Europe, accessed November 30, 2025, https://nebius.com/newsroom/nebius-to-invest-more-than-usd-1-billion-to-build-ai-infrastructure-in-europe Is Nebius a Buy? | The Motley Fool, accessed November 30, 2025, https://www.fool.com/investing/2025/11/28/is-nebius-a-buy/ Applied Digital to classify Cloud Services as discontinued operations - Investing.com, accessed November 30, 2025, https://www.investing.com/news/sec-filings/applied-digital-to-classify-cloud-services-as-discontinued-operations-93CH-4073767 Applied Digital Completes Phase II Ready for Service at Polaris Forge 1, Fully Energizing First 100 MW Building, accessed November 30, 2025, https://ir.applieddigital.com/news-events/press-releases/detail/137/applied-digital-completes-phase-ii-ready-for-service-at Applied Digital Reports Fiscal First Quarter 2026 Results, accessed November 30, 2025, https://ir.applieddigital.com/news-events/press-releases/detail/131/applied-digital-reports-fiscal-first-quarter-2026-results IREN Deep Dive: Profitable Bitcoin Mining Powering AI Infrastructure, accessed November 30, 2025, https://mlq.ai/research/iren-deep-dive-profitable-bitcoin-mining-powering-ai-infrastructure/ IREN Secures New Multi-Year AI Cloud Contracts, accessed November 30, 2025, https://irisenergy.gcs-web.com/news-releases/news-release-details/iren-secures-new-multi-year-ai-cloud-contracts Compute | Hut 8, accessed November 30, 2025, https://www.hut8.com/compute What is Sovereign Cloud? - IBM, accessed November 30, 2025, https://www.ibm.com/think/topics/sovereign-cloud What is Sovereign Public Cloud - Microsoft Learn, accessed November 30, 2025, https://learn.microsoft.com/en-us/industry/sovereign-cloud/sovereign-public-cloud/overview-sovereign-public-cloud Oracle and Palantir Join Forces to Deliver Mission Critical AI Solutions to Governments and Businesses, accessed November 30, 2025, https://www.oracle.com/news/announcement/oracle-and-palantir-join-forces-to-deliver-mission-critical-ai-solutions-to-governments-and-businesses-2024-04-04/ NATO signs major Google deal to ensure secure sovereign cloud, accessed November 30, 2025, https://www.techradar.com/pro/nato-signs-major-google-deal-to-ensure-secure-sovereign-cloud Datadog stock rockets 20% on AI-powered revenue beat | The Tech Buzz, accessed November 30, 2025, https://www.techbuzz.ai/articles/datadog-stock-rockets-20-on-ai-powered-revenue-beat Arista Delivers Holistic AI Solutions in Collaboration with NVIDIA, accessed November 30, 2025, https://investors.arista.com/Communications/Press-Releases-and-Events/Press-Release-Detail/2024/Arista-Delivers-Holistic-AI-Solutions-in-Collaboration-with-NVIDIA/default.aspx AI Networking - Arista, accessed November 30, 2025, https://www.arista.com/assets/data/pdf/Whitepapers/AI-Network-WP.pdf Pure Storage Announces Strategic Investment and Technology Partnership with CoreWeave to Accelerate Large-Scale AI Cloud Services Innovation - PR Newswire, accessed November 30, 2025, https://www.prnewswire.com/news-releases/pure-storage-announces-strategic-investment-and-technology-partnership-with-coreweave-to-accelerate-large-scale-ai-cloud-services-innovation-302303548.html Digi Power X Acquires Supermicro NVIDIA B200 Systems to Launch Tier 3 NeoCloud AI Infrastructure | Nasdaq, accessed November 30, 2025, https://www.nasdaq.com/press-release/digi-power-x-acquires-supermicro-nvidia-b200-systems-launch-tier-3-neocloud-ai
You’re currently a free subscriber to 大叔美股筆記 Uncle Stock Notes. For the full experience, upgrade your subscription. 您目前是大叔美股筆記 Uncle Stock Notes 的免費訂閱者。要獲得完整體驗，請升級您的訂閱。 本篇內容為大叔個人投資筆記與心得分享，僅供參考，並不構成任何投資建議或買賣操作的依據。
投資一定有風險，有賺也有賠，入市前請務必自行評估風險承受度與財務狀況，做好功課、審慎決策，才是對自己最好的保護。 謝謝大家的理解與支持，祝大家都能穩穩獲利，長長久久！ #投資有風險 #請審慎評估 #大叔只是分享不是報明牌 This content represents my personal investment notes and experience only, and does not constitute financial or investment advice.
Please be aware that all investments involve risks and may result in gains or losses. Before making any investment decisions, carefully evaluate your risk tolerance and financial situation. Do your research and make prudent choices to protect yourself. Thank you for your understanding and support. Wishing everyone steady gains and long-term success! #InvestAtYourOwnRisk #NotFinancialAdvice #UncleSharingNotStockTip | |
No comments:
Post a Comment